From: Raphael Hoffmann (raphaelh_at_u.washington.edu)
Date: Sun Nov 23 2003 - 19:26:41 PST
Review: "Acting Optimally in Partially Observable Stochastic Domains" by
Anthony R. Cassandra, Leslie Pack Kaelbling and Michael L. Littman
Solving POMDPs is a computationally challenging task. This paper presents a
new, (relatively) fast algorithm that approximates an optimal solution.
Compared to the previous papers we read, this one is extremely readable. The
structure of the text is very clear and every step in the argumentation is
explained very well. Rather tahtn using lots of mathematical definitions and
transformations, the authors write in a clear and understandable language,
focusing on the relevant parts.
Furthermore, several simple examples with figures help to understand the
ideas. E.g. the tiger example definitly shows the importance of "Acting to
gain information".
Although I liked the shortness of the paper, some important details are
obviously missing. The authors claim that their algorithm is by far more
efficient than precious ones, without actually presenting detailed
comparisons. It is not clear how much faster their algorithm really is. In
fact, the whole chapter "Results" does not contain any numbers.
They only briefly mentioned that a problem of 24 states, 4 actions and 11
observations took them about half an hour to compute. To me that still
appears to be too much for many real-world problems where the state space
could be significantly larger and where responses are often required in
real-time.
Certainly, a precise and detailed comparison to similar algorithms is
missing. That should be done as future work.
In the text, the authors also briefly mentioned that the Witness algorithm
becomes numerically unstable if set to low delta, i.e. low deviation from
optimal solution. I belive that this fact should be analysed more
thoroughly. (What are maximum numerical errors? Can these effects be
reduced?).
This archive was generated by hypermail 2.1.6 : Sun Nov 23 2003 - 19:26:44 PST